Downloads · 30 days
14
25% of all-time downloads
leslie721007/babylm-strict-small-coherent86-alpha075
babylm-strict-small-coherent86-alpha075 is a fill-mask model from leslie721007. Use it when you need the model to fill a missing word. It is set up for transformers. The card lists the license as other.
Repository target: leslie721007/babylm-strict-small-coherent86-alpha075 Revision hint at construction time: main
Downloads · 30 days
14
25% of all-time downloads
All-time downloads
57
Public
Parameters
36.5M
146 MB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors146 MB · 97%
From the Hugging Face model README
Repository target: leslie721007/babylm-strict-small-coherent86-alpha075
Revision hint at construction time: main
This repository contains the exact coherent86 alpha=0.75 masked-language-model
function from session S0829A02. It is the current highest local endpoint
candidate in the session record, but it is not a complete BabyLM submission
package. The real AoA and fast-checkpoint materials are being completed and
independently verified in the paired session; this repository must not be used
as a substitute for that complete package.
Use trusted code:
from transformers import AutoModelForMaskedLM, AutoTokenizer
repo = "leslie721007/babylm-strict-small-coherent86-alpha075"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForMaskedLM.from_pretrained(repo, trust_remote_code=True)
Without trust_remote_code=True, Transformers can instantiate a native
DebertaV2ForMaskedLM and ignore the private-path tensors. That native fallback
has fewer parameters and is not this scored function.
FrozenSlowPrivateDebertaV2ForMaskedLMprivate_adapter_enabled: trueprivate_adapter_scale: 0.7536,458,59235,463,008995,584model.safetensors SHA256: e14d757ae51b41e33bf0813f841248fecd1eefeb9e040f520c4c6203343b15c8config.json SHA256: ca6792fe1842f0e89e7080e7709e0619fdcaf5ce48a65f69bf7137ac26ce9bb7tokenizer.json SHA256: a9cbb830495cb92bbb2996adc256207746282ee67f4f8d40a4f646a634ec139a82,012,495 counted words3,992,800 counted words86,005,295 counted wordsThe alpha=0.75 function shares the same learned coherent86 private-path weights
as alpha=1.0 but has a different config.json; the scale value is copied into
the module at construction. Changing a loaded config field after construction is
not enough to obtain this function.
Eight official-compatible measured columns are available for this exact alpha=0.75 function:
| measured column | score |
|---|---|
| BLiMP | 68.51 |
| Supplement | 63.64 |
| EWoK | 50.02 |
| Entity | 28.32 |
| COMPS | 52.05 |
| GlobalPIQA | 38.565 |
| Reading | 8.165 |
| SuperGLUE | 69.8192223897 |
Cheap7 from the seven non-SuperGLUE columns is 44.181428571429;
SuperGLUE is 69.819222389699. The arithmetic value
42.121024709967 assumes AoA=0 and is
recorded only as a provisional local arithmetic coordinate. The final submission
package requires the real AoA and fast-checkpoint materials; those are not
filled in here.
The protected public fallback remains the previously submitted chck_82M
endpoint at displayed 41.94. Coherent86 alpha=0.75 is a stronger local endpoint
function, but session analyses read the private-scale family as amplitude-driven
competence redistribution rather than a settled general slow-fast learning
mechanism. This repository is therefore an exact endpoint carrier for human and
paired-session verification, not a final leaderboard submission.